Improved satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis and detection method thereof

Abstract

The invention discloses an improved satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis and a detection method thereof. The system comprises a time-frequency transformation module, a time-frequency image processing module and a neural network classification module. The time-frequency transformation module comprises time-frequency transformation modes of short-time Fourier transformation, continuous wavelet transformation and Margenauu-Hill transformation, the time-frequency transformation module directly takes an intermediate-frequency digital signal as input, and the defect that the input signals are different due to different front ends is overcome; the time-frequency image processing module is composed of a time-frequency standardization module, a gray processing module and a multi-channel image fusion module. The multi-channel image fusion module fuses three types of time-frequency transformation characteristics, different time-frequency transformation methods show different performance expressions under different signal characteristics and environment conditions, and a multi-level time-frequency analysis method is fused to endow the system with higher signal intelligent adaptability, so that the system is excellent in performance when facing various signal types and complex scenes, and the system is more suitable for various signal types and complex scenes. And the anti-noise performance is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of GNSS interference detection, and in particular relates to an improved satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis and a detection method thereof. Background Art

[0002] The Global Navigation Satellite System (GNSS) is widely used in various fields of modern society. The location services it provides have brought people a good experience. However, due to the lack of security protection and vulnerability of navigation signals, satellite navigation spoofing interference is considered to be the main threat to life safety applications. Therefore, enhancing the anti-spoofing capability of the global navigation satellite system is an issue that urgently needs to be studied and solved. It is one of the key technologies to ensure the accuracy and reliability of global satellite navigation information.

[0003] The forwarding deception interference is to receive the real satellite signal through the receiving antenna, then add a certain delay through the delay device, and amplify the signal power, broadcast the deception signal through the transmitting antenna, control its code phase to be several chips outside the target receiver tracking signal, and then make the generated deception interference signal reach the target receiver antenna phase center and align with the real signal code phase, and use the power advantage of the deception signal to make the target receiver track the deception signal. According to the above characteristics of the deception signal, the detection method that is currently widely used is to pre-process the extracted feature quantities such as Ratio, Delta and carrier-to-noise ratio, and input the processed feature quantities into the BP neural network of the appropriate architecture to obtain the detection result of the deception interference signal. However, the input parameters of the BP neural network should be relatively independent, and a large number of detection quantity correlation analyses are required to obtain the appropriate detection quantity input, making the pre-processing work complicated and cumbersome. The present invention performs time-frequency analysis on the deception signal. Compared with the BP neural network feature quantity extraction method, the time-frequency analysis method directly analyzes the intermediate frequency digital signal, which not only simplifies the pre-processing process to a certain extent, but also provides the joint distribution information of the time domain and the frequency domain, thereby accurately describing the relationship between the change of the signal frequency over time. In addition, the BP neural network analysis method based on signal feature extraction has the disadvantages of low accuracy and low efficiency due to the limitations of the extracted signal feature types.

[0004] Therefore, the present invention analyzes the time-frequency characteristics of the signal through a convolutional neural network. First, the collected intermediate frequency signal is transformed in different time-frequency domains. After the normalized preprocessed image is input into the convolutional neural network, the image features are extracted after multiple convolution pooling operations. Finally, the image features are sent to a fully connected network to complete the detection of deceptive signals. Summary of the invention

[0005] In view of the complexity and tediousness of signal feature extraction and the difficulty of capturing local image features, the present invention proposes an improved satellite navigation deception detection system and detection method based on deep learning fusion multi-level time-frequency analysis.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] An improved satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis includes a time-frequency transformation module, a time-frequency image processing module and a neural network classification module; wherein the time-frequency transformation module includes three time-frequency transformation methods, namely short-time Fourier transform (STFT), continuous wavelet transform (CWT) and Margenau-Hill transform (MH), and the time-frequency transformation module directly uses an intermediate frequency digital signal as input, thereby overcoming the defect that input signals are different due to different front ends; the time-frequency image processing module includes a time-frequency normalization module, a grayscale processing module and a multi-channel image fusion module; the multi-channel image fusion module fuses three types of time-frequency transformation features to the neural network classification module, and the output end of the neural network classification module outputs a deception interference signal detection result; wherein the multi-channel image fusion module fuses three types of time-frequency transformation features, and different time-frequency transformation methods present different performances under different signal characteristics and environmental conditions. The fusion of multi-level time-frequency analysis methods gives the system stronger signal intelligence adaptability, so that it performs excellently in the face of various signal types and complex scenes, the analysis is more extensive, and the anti-noise performance is improved.

[0008] Furthermore, the connection structure of each module of the interference signal detection is as follows: the GNSS intermediate frequency digital signal collected by the front end is used as the input end of the time-frequency conversion module, the output end of the time-frequency conversion module is connected to the input end of the time-frequency image processing module, the output end of the time-frequency image processing module is connected to the input end of the neural network classification module, and the output end of the neural network classification module outputs the deception interference signal detection result;

[0009] Furthermore, the connection structure of each module inside the time-frequency image processing module: the time-frequency normalization module and the normalization processing module respectively include three input terminals and three output terminals, the three output terminals of the time-frequency normalization module are correspondingly connected to the three input terminals of the normalization processing module, the three output terminals of the normalization processing module are connected to the input terminal of the multi-channel image fusion module, and the output terminal of the multi-channel image fusion module serves as the output terminal of the time-frequency image processing module.

[0010] An improved satellite navigation deception detection method based on deep learning fusion multi-level time-frequency analysis includes the following steps:

[0011] S1, input the collected and processed intermediate frequency digital signal into the time-frequency transformation module for time-frequency analysis, respectively use short-time Fourier transform, continuous wavelet transform and Margenau-Hill transform to analyze the intermediate frequency digital signal, obtain the joint distribution information of the signal in the time domain and frequency domain, and output the time-frequency spectrum;

[0012] S2. For the time-frequency spectrograms output by the three types of transformations, the time-frequency normalization module performs time-frequency normalization processing on them respectively, and uses the bicubic difference method to adjust the size of the image to the size of the corresponding neural network to 227×227;

[0013] S3, input the normalized three types of images into the normalization processing module for normalization processing respectively, and process the image data into grayscale data of 0 to 255 through linear function normalization;

[0014] S4, the three types of time-frequency transformed images processed by steps S1-S3 are input into a multi-channel image fusion module to fuse the three types of two-dimensional time-frequency images of the same signal into a 227×227×3 three-channel image;

[0015] S5. Input the processed image into the neural network classification module. The input layer has the same size and dimension as the image. The output layer contains 2 nodes, which is the same as the number of output classification targets. By designing the number of hidden layers and nodes, and training with simulation and semi-physical simulation data, the detection of deceptive interference signals can be finally achieved.

[0016] Furthermore, in step S1, the short-time Fourier transform is:

[0017] For a signal s(p) with limited energy, its short-time Fourier transform is defined as:

[0018]

[0019] Where: s(p) is the input signal, p is the index of the sampling point in the signal sequence, h(pm) is the moving window function, which indicates the offset of the window function relative to the central time point m, m is the central position of the window function, f is the frequency of the signal, F s is the sampling frequency, N is the length of the window function, e is the base of the natural logarithm, and j is the imaginary unit; further, the time-frequency energy density spectrum q(f,m) is:

[0020]

[0021] Where A is a real-valued scalar. For the unilateral energy spectral density, the calculation formula is as follows:

[0022]

[0023] Where: h(p) represents a fixed window function, and at the zero frequency and the Nyquist frequency, the factor 2 in the numerator is changed to 1.

[0024] Furthermore, in step S1, continuous wavelet transform:

[0025] For a signal s(p) with limited energy, its wavelet transform is defined as:

[0026]

[0027] Among them, WT s (a, b) Wavelet coefficients, representing the projection intensity of the signal at scale a and time position b; is the basic wavelet; a, b are constants, representing The factor for scaling and time shifting, and a>0; * represents the complex conjugate, dp represents the infinitesimal component of p;

[0028] Wavelet transform selects different time windows for different frequencies, and selects narrower time windows for high-frequency signals, whereas wider time windows are selected for low-frequency signals. One advantage of wavelet transform over Fourier transform is that it can display the local characteristics of the signal, while Fourier transform studies the overall characteristics of the signal.

[0029] Furthermore, in step S1, Margenau-Hill transformation:

[0030] The Margenau-Hill transform is essentially a bilinear time-frequency distribution, which defines an instantaneous correlation function for the signal, which is expressed as:

[0031]

[0032] Among them, P MH represents the Margenau-Hill distribution, is the kernel function of windowed smoothing, u represents the time of time-frequency transformation, τ represents the time shift, ξ is the frequency shift parameter, and ω is the angular frequency of the signal;

[0033] The Margenau-Hill transform has high time-frequency resolution and good time-frequency aggregation. Through windowing and smoothing of cos(πτv), it overcomes the cross-interference caused by the interaction between signals when facing multi-component signals or signals with complex modulation rules, thereby more accurately restoring the original characteristics of the signal.

[0034] Furthermore, in step S2, time-frequency normalization is performed:

[0035] The three types of time-frequency transformed images were normalized, and the images were enhanced using bicubic interpolation. The image size was adjusted to 227×227.

[0036]

[0037] Among them, (y,z) represents the coordinates of the input sample feature point, f(y,z) represents the value that should be inserted after calculation of the pixel to be interpolated, (y α ,z β ) represents the points in the 4x4 area near the input sample feature point, where: α, β = 0, 1, 2, 3, and the W function is called the BiCubic basis function:

[0038]

[0039] Where: W(k) is the BiCubic basis function, |k| represents the distance between the target point and the surrounding points, and λ is the smoothness parameter;

[0040] After calculation, it is concluded that λ = -1 / 2, in order to achieve and provide the best interpolation quality:

[0041]

[0042] Cubic convolution interpolation uses the color values ​​of 16 points around the sampling point, combined with other factors, such as the color change rate of adjacent points, and other factors that affect the interpolation result, to achieve a more accurate interpolation effect, making the adjusted image resolution closer to the original image.

[0043] Furthermore, in step S3, the normalization process is performed:

[0044] The normalized time-frequency image is normalized by a linear function. The normalized time-frequency image is represented as S(t,ω), where t and ω represent the time and frequency corresponding to the input signal respectively. The normalized image information is represented as:

[0045]

[0046] Where: S norm (t,ω) is the normalized data, S(t,ω) is the original data, S max (t,ω),S min (t,ω) are the maximum and minimum values ​​of the original data set respectively.

[0047] Furthermore, in step S4, multi-channel image fusion:

[0048] The normalized images of the three types of time-frequency transforms are fused in multi-channel. The normalized images are grayscale images with pixel values ​​of 0 to 255. The images of short-time Fourier transform, continuous wavelet transform and Margenau-Hill transform results after time-frequency normalization and normalization are fused in multi-channel respectively; the fused image is expressed as:

[0049]

[0050] Where: L is the depth of the three-channel image and They are the images of the signal after STFT, wavelet transform and Margenau-Hill transform respectively.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] (1) The present invention proposes a time-frequency analysis method to extract the features of the spoofing signal. Compared with the traditional signal feature extraction method of Fourier transform, it provides more accurate time domain information and richer frequency domain features, and obtains more comprehensive joint distribution information;

[0053] (2) The present invention adopts a convolutional neural network to identify time-frequency feature images. Compared with the traditional analysis method based on signal feature extraction, it overcomes the limitations of different feature signals and improves accuracy and efficiency;

[0054] (3) The present invention adopts three different types of time-frequency transformation methods, has the ability of multi-scale analysis, can capture the global and local characteristics of the signal at the same time, and increases the robustness of signal feature extraction;

[0055] (4) The present invention integrates the characteristics of three types of time-frequency transformation through multi-channel image fusion, obtains multi-directional joint time-frequency distribution information of the signal, and improves the accuracy of image recognition;

[0056] (5) The present invention is based on software algorithm and has the advantages of great flexibility and low cost.

[0057] (6) The technology of the present invention analyzes intermediate frequency signals and is not affected by the satellite signal structure. It is suitable for detecting deception interference signals of various GNSS systems such as GPS and Beidou. It has strong versatility and important engineering significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the overall structural block diagram of the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the embodiments.

[0060] Example 1

[0061] like Figure 1 As shown, an improved satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis includes a time-frequency transformation module, a time-frequency image processing module and a neural network classification module; wherein the time-frequency transformation module includes three time-frequency transformation methods, namely short-time Fourier transform (STFT), continuous wavelet transform (CWT) and Margenau-Hill transform (MH), and the time-frequency transformation module directly uses the intermediate frequency digital signal as input, thereby overcoming the defect that the input signal has differences due to different front ends; the time-frequency image processing module includes a time-frequency normalization module, a grayscale processing module and a multi-channel image fusion module; the multi-channel image fusion module fuses three types of time-frequency transformation features to the neural network classification module, and the output end of the neural network classification module outputs the deception interference signal detection result; wherein the multi-channel image fusion module fuses three types of time-frequency transformation features, and different time-frequency transformation methods present different performances under different signal characteristics and environmental conditions. The fusion of multi-level time-frequency analysis methods gives the system stronger signal intelligence adaptability, so that it performs excellently in the face of various signal types and complex scenes, the analysis is more extensive, and the anti-noise performance is improved.

[0062] Connection structure of each module of interference signal detection: the GNSS intermediate frequency digital signal collected by the front end is used as the input of the time-frequency conversion module, the output of the time-frequency conversion module is connected to the input of the time-frequency image processing module, the output of the time-frequency image processing module is connected to the input of the neural network classification module, and the output of the neural network classification module outputs the deception interference signal detection result;

[0063] The connection structure of each module inside the time-frequency image processing module: the time-frequency normalization module and the normalization processing module respectively contain three input terminals and three output terminals. The three output terminals of the time-frequency normalization module are correspondingly connected to the three input terminals of the normalization processing module. The three output terminals of the normalization processing module are connected to the input terminal of the multi-channel image fusion module. The output terminal of the multi-channel image fusion module serves as the output terminal of the time-frequency image processing module.

[0064] Example 2

[0065] An improved satellite navigation deception detection method based on deep learning fusion multi-level time-frequency analysis includes the following steps:

[0066] S1, input the collected and processed intermediate frequency digital signal into the time-frequency transformation module for time-frequency analysis, respectively use short-time Fourier transform, continuous wavelet transform and Margenau-Hill transform to analyze the intermediate frequency digital signal, obtain the joint distribution information of the signal in the time domain and frequency domain, and output the time-frequency spectrum;

[0067] In step S1, short-time Fourier transform:

[0068] For a signal s(p) with limited energy, its short-time Fourier transform is defined as:

[0069]

[0070] Where: s(p) is the input signal, p is the index of the sampling point in the signal sequence, h(pm) is the moving window function, which indicates the offset of the window function relative to the central time point m, m is the central position of the window function, f is the frequency of the signal, F s is the sampling frequency, N is the length of the window function, e is the base of the natural logarithm, and j is the imaginary unit; further, the time-frequency energy density spectrum q(f,m) is:

[0071]

[0072] Where A is a real-valued scalar. For the unilateral energy spectral density, the calculation formula is as follows:

[0073]

[0074] Where: h(p) represents a fixed window function, and at the zero frequency and the Nyquist frequency, the factor 2 in the numerator is changed to 1.

[0075] In step S1, continuous wavelet transform:

[0076] For a signal s(p) with limited energy, its wavelet transform is defined as:

[0077]

[0078] Among them, WT s (a, b) Wavelet coefficients, representing the projection intensity of the signal at scale a and time position b; is the basic wavelet; a, b are constants, representing The factor for scaling and time shifting, and a>0; * represents the complex conjugate, dp represents the infinitesimal component of p;

[0079] Wavelet transform selects different time windows for different frequencies, and selects narrower time windows for high-frequency signals, whereas wider time windows are selected for low-frequency signals. One advantage of wavelet transform over Fourier transform is that it can display the local characteristics of the signal, while Fourier transform studies the overall characteristics of the signal.

[0080] In step S1, Margenau-Hill transformation:

[0081] The Margenau-Hill transform is essentially a bilinear time-frequency distribution, which defines an instantaneous correlation function for the signal, which is expressed as:

[0082]

[0083] Among them, P MH represents the Margenau-Hill distribution, is the kernel function of windowed smoothing, u represents the time of time-frequency transformation, τ represents the time shift, ξ is the frequency shift parameter, and ω is the angular frequency of the signal;

[0084] The Margenau-Hill transform has high time-frequency resolution and good time-frequency aggregation. Through windowing and smoothing of cos(πτv), it overcomes the cross-interference caused by the interaction between signals when facing multi-component signals or signals with complex modulation rules, thereby more accurately restoring the original characteristics of the signal.

[0085] S2. For the time-frequency spectrograms output by the three types of transformations, the time-frequency normalization module performs time-frequency normalization processing on them respectively, and uses the bicubic difference method to adjust the size of the image to the size of the corresponding neural network to 227×227;

[0086] In step S2, time-frequency normalization is performed:

[0087] The three types of time-frequency transformed images were normalized, and the images were enhanced using bicubic interpolation. The image size was adjusted to 227×227.

[0088]

[0089] Among them, (y,z) represents the coordinates of the input sample feature point, f(y,z) represents the value that should be inserted after calculation of the pixel to be interpolated, (y α ,z β ) represents the points in the 4x4 area near the input sample feature point, where: α, β = 0, 1, 2, 3, and the W function is called the BiCubic basis function:

[0090]

[0091] Where: W(k) is the BiCubic basis function, |k| represents the distance between the target point and the surrounding points, and λ is the smoothness parameter;

[0092] After calculation, it is concluded that λ = -1 / 2, in order to achieve and provide the best interpolation quality:

[0093]

[0094] Cubic convolution interpolation uses the color values ​​of 16 points around the sampling point, combined with other factors, such as the color change rate of adjacent points, and other factors that affect the interpolation result, to achieve a more accurate interpolation effect, making the adjusted image resolution closer to the original image.

[0095] S3, input the normalized three types of images into the normalization processing module for normalization processing respectively, and process the image data into grayscale data of 0 to 255 through linear function normalization;

[0096] In step S3, the normalization process is:

[0097] The normalized time-frequency image is normalized by a linear function. The normalized time-frequency image is represented as S(t,ω), where t and ω represent the time and frequency corresponding to the input signal respectively. The normalized image information is represented as:

[0098]

[0099] Where: S norm (t,ω) is the normalized data, S(t,ω) is the original data, S max (t,ω),S min (t,ω) are the maximum and minimum values ​​of the original data set respectively.

[0100] S4, the three types of time-frequency transformed images processed by steps S1-S3 are input into a multi-channel image fusion module to fuse the three types of two-dimensional time-frequency images of the same signal into a 227×227×3 three-channel image;

[0101] In step S4, multi-channel image fusion:

[0102] The normalized images of the three types of time-frequency transforms are fused in multi-channel. The normalized images are grayscale images with pixel values ​​of 0 to 255. The images of short-time Fourier transform, continuous wavelet transform and Margenau-Hill transform results after time-frequency normalization and normalization are fused in multi-channel respectively; the fused image is expressed as:

[0103]

[0104] Where: L is the depth of the three-channel image and They are the images of the signal after STFT, wavelet transform and Margenau-Hill transform respectively.

[0105] S5. Input the processed image into the neural network classification module. The input layer has the same size and dimension as the image. The output layer contains 2 nodes, which is the same as the number of output classification targets. By designing the number of hidden layers and nodes, and training with simulation and semi-physical simulation data, the detection of deceptive interference signals can be finally achieved.

[0106] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An improved satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis, characterized in that: It includes a time-frequency transformation module, a time-frequency image processing module and a neural network classification module; the time-frequency transformation module includes three time-frequency transformation methods, namely short-time Fourier transform, continuous wavelet transform and Margenau-Hill transform, and the time-frequency transformation module directly uses the intermediate frequency digital signal as input; the time-frequency image processing module includes a time-frequency normalization module, a grayscale processing module and a multi-channel image fusion module; the multi-channel image fusion module fuses three types of time-frequency transformation features to the neural network classification module, and the output end of the neural network classification module outputs the deception interference signal detection result.

2. According to claim 1, an improved satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis is characterized in that: The GNSS intermediate frequency digital signal collected by the front end is used as the input end of the time-frequency conversion module, the output end of the time-frequency conversion module is connected to the input end of the time-frequency image processing module, the output end of the time-frequency image processing module is connected to the input end of the neural network classification module, and the output end of the neural network classification module outputs the deception interference signal detection result; The time-frequency normalization module and the normalization processing module respectively include three input terminals and three output terminals. The three output terminals of the time-frequency normalization module are correspondingly connected to the three input terminals of the normalization processing module. The three output terminals of the normalization processing module are connected to the input terminal of the multi-channel image fusion module. The output terminal of the multi-channel image fusion module serves as the output terminal of the time-frequency image processing module.

3. A detection method for improving a satellite navigation deception detection system based on deep learning fusion multi-level time-frequency analysis according to any one of claims 1-2, characterized in that: The following steps are involved: S1, input the collected and processed intermediate frequency digital signal into the time-frequency transformation module for time-frequency analysis, respectively use short-time Fourier transform, continuous wavelet transform and Margenau-Hill transform to analyze the intermediate frequency digital signal, obtain the joint distribution information of the signal in the time domain and frequency domain, and output the time-frequency spectrum; S2. For the time-frequency spectrograms output by the three types of transformations, the time-frequency normalization module performs time-frequency normalization processing on them respectively, and uses the bicubic difference method to adjust the size of the image to the size of the corresponding neural network to 227×227; S3, input the normalized three types of images into the normalization processing module for normalization processing respectively, and process the image data into grayscale data of 0 to 255 through linear function normalization; S4, the three types of time-frequency transformed images processed by steps S1-S3 are input into a multi-channel image fusion module to fuse the three types of two-dimensional time-frequency images of the same signal into a 227×227×3 three-channel image; S5. Input the processed image into the neural network classification module. The input layer has the same size and dimension as the image. The output layer contains 2 nodes, which is the same as the number of output classification targets. By designing the number of hidden layers and nodes, and training with simulation and semi-physical simulation data, the detection of deceptive interference signals can be finally achieved.

4. According to claim 3, the improved satellite navigation deception detection method based on deep learning fusion multi-level time-frequency analysis is characterized in that: In step S1, short-time Fourier transform: For a signal s(p) with limited energy, its short-time Fourier transform is defined as: Where: s(p) is the input signal, p is the index of the sampling point in the signal sequence, h(pm) is the moving window function, which indicates the offset of the window function relative to the central time point m, m is the central position of the window function, f is the frequency of the signal, F s is the sampling frequency, N is the length of the window function, e is the base of the natural logarithm, and j is the imaginary unit; further, the time-frequency energy density spectrum q(f,m) is: Where A is a real-valued scalar. For the unilateral energy spectral density, the calculation formula is as follows: Where: h(p) represents a fixed window function, and at the zero frequency and the Nyquist frequency, the factor 2 in the numerator is changed to 1.

5. According to claim 3, the improved satellite navigation deception detection method based on deep learning fusion multi-level time-frequency analysis is characterized in that: In step S1, continuous wavelet transform: For a signal s(p) with limited energy, its wavelet transform is defined as: Among them, WT s (a, b) Wavelet coefficients, representing the projection intensity of the signal at scale a and time position b; is the basic wavelet; a, b are constants, representing The factor for scaling and time shifting, and a>0; * represents the complex conjugate, dp represents the infinitesimal component of p; Wavelet transform selects different time windows for different frequencies, and selects narrower time windows for high-frequency signals, whereas wider time windows are selected for low-frequency signals. One advantage of wavelet transform over Fourier transform is that it can display the local characteristics of the signal, while Fourier transform studies the overall characteristics of the signal.

6. According to claim 3, the improved satellite navigation deception detection method based on deep learning fusion multi-level time-frequency analysis is characterized in that: In step S1, Margenau-Hill transformation: The Margenau-Hill transform is essentially a bilinear time-frequency distribution, which defines an instantaneous correlation function for the signal, which is expressed as: Among them, P MH represents the Margenau-Hill distribution, is the kernel function of windowed smoothing, u represents the time of time-frequency transformation, τ represents the time shift, ξ is the frequency shift parameter, and ω is the angular frequency of the signal; The Margenau-Hill transform has high time-frequency resolution and good time-frequency aggregation. Through windowing and smoothing of cos(πτv), it overcomes the cross-interference caused by the interaction between signals when facing multi-component signals or signals with complex modulation rules, thereby more accurately restoring the original characteristics of the signal.

7. The improved satellite navigation deception detection method based on deep learning fusion multi-level time-frequency analysis according to claim 3 is characterized in that: In step S2, time-frequency normalization is performed: The three types of time-frequency transformed images were normalized, and the images were enhanced using bicubic interpolation. The image size was adjusted to 227×227. Among them, (y,z) represents the coordinates of the input sample feature point, f(y,z) represents the value that should be inserted after calculation of the pixel to be interpolated, (y α ,z β ) represents the points in the 4x4 area near the input sample feature point, where: α, β = 0, 1, 2, 3, and the W function is called the BiCubic basis function: Where: W(k) is the BiCubic basis function, |k| represents the distance between the target point and the surrounding points, and λ is the smoothness parameter; After calculation, it is concluded that λ = -1 / 2, in order to achieve and provide the best interpolation quality: Cubic convolution interpolation uses the color values ​​of 16 points around the sampling point, combined with other factors, such as the color change rate of adjacent points, and other factors that affect the interpolation result, to achieve a more accurate interpolation effect, making the adjusted image resolution closer to the original image.

8. The improved satellite navigation spoofing detection method based on deep learning fusion multi-level time-frequency analysis according to claim 3 is characterized in that: In step S3, the normalization process is: The normalized time-frequency image is normalized by a linear function. The normalized time-frequency image is represented as S(t,ω), where t and ω represent the time and frequency corresponding to the input signal respectively. The normalized image information is represented as: Where: S norm (t,ω) is the normalized data, S(t,ω) is the original data, S max (t,ω),S min (t,ω) are the maximum and minimum values ​​of the original data set respectively.

9. The improved satellite navigation deception detection method based on deep learning fusion multi-level time-frequency analysis according to claim 3 is characterized in that: In step S4, multi-channel image fusion: The normalized images of the three types of time-frequency transforms are fused in multi-channel. The normalized images are in grayscale format with pixel values ​​of 0 to 255. The images of short-time Fourier transform, continuous wavelet transform and Margenau-Hill transform results after time-frequency normalization and normalization are fused in multi-channel respectively; the fused image is expressed as: Where: L is the depth of the three-channel image and They are the images of the signal after STFT, wavelet transform and Margenau-Hill transform respectively.

Images